1. 18-447 Computer Architecture Lecture 7: Pipelining Prof. Onur Mutlu Carnegie Mellon University Spring 2014, 1/29/2014

2. Can We Do Better than Microprogrammed Designs? What limitations do you see with the multi-cycle design? Limited concurrency  Some hardware resources are idle during different phases of instruction processing cycle  “Fetch” logic is idle when an instruction is being “decoded” or “executed”  Most of the datapath is idle when a memory access is happening 2

3. Can We Use the Idle Hardware to Improve Concurrency? Goal: Concurrency  throughput (more “work” completed in one cycle) Idea: When an instruction is using some resources in its processing phase, process other instructions on idle resources not needed by that instruction  E.g., when an instruction is being decoded, fetch the next instruction  E.g., when an instruction is being executed, decode another instruction  E.g., when an instruction is accessing data memory (ld/st), execute the next instruction  E.g., when an instruction is writing its result into the register file, access data memory for the next instruction 3

4. Pipelining: Basic Idea More systematically:  Pipeline the execution of multiple instructions  Analogy: “Assembly line processing” of instructions Idea:  Divide the instruction processing cycle into distinct “stages” of processing  Ensure there are enough hardware resources to process one instruction in each stage  Process a different instruction in each stage Instructions consecutive in program order are processed in consecutive stages Benefit: Increases instruction processing throughput (1/CPI) Downside: Start thinking about this… 4

5. Example: Execution of Four Independent ADDs Multi-cycle: 4 cycles per instruction Pipelined: 4 cycles per 4 instructions (steady state) 5 Time FDEWFDEWFDEWFDEWFDEWFDEWFDEWFDEW Is life always this beautiful?

6. The Laundry Analogy “place one dirty load of clothes in the washer” “when the washer is finished, place the wet load in the dryer” “when the dryer is finished, take out the dry load and fold” “when folding is finished, ask your roommate (??) to put the clothes away” 6 - steps to do a load are sequentially dependent - no dependence between different loads - different steps do not share resources Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.]

7. Pipelining Multiple Loads of Laundry 7 - latency per load is the same - throughput increased by 4 - 4 loads of laundry in parallel - no additional resources Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.]

8. Pipelining Multiple Loads of Laundry: In Practice 8 the slowest step decides throughput Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.]

9. Pipelining Multiple Loads of Laundry: In Practice 9 A B A B Throughput restored (2 loads per hour) using 2 dryers Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.]

10. An Ideal Pipeline Goal: Increase throughput with little increase in cost (hardware cost, in case of instruction processing) Repetition of identical operations  The same operation is repeated on a large number of different inputs Repetition of independent operations  No dependencies between repeated operations Uniformly partitionable suboperations  Processing can be evenly divided into uniform-latency suboperations (that do not share resources) Fitting examples: automobile assembly line, doing laundry  What about the instruction processing “cycle”? 10

11. Ideal Pipelining 11 combinational logic (F,D,E,M,W) T psec BW=~(1/T) BW=~(2/T) T/2 ps (F,D,E)T/2 ps (M,W) BW=~(3/T) T/3 ps (F,D) T/3 ps (E,M) T/3 ps (M,W)

12. More Realistic Pipeline: Throughput Nonpipelined version with delay T BW = 1/(T+S) where S = latch delay k-stage pipelined version BW k-stage = 1 / (T/k +S ) BW max = 1 / (1 gate delay + S ) 12 T ps T/k ps T/k ps

13. More Realistic Pipeline: Cost Nonpipelined version with combinational cost G Cost = G+L where L = latch cost k-stage pipelined version Cost k-stage = G + Lk 13 G gates G/k

14. Pipelining Instruction Processing 14

15. Remember: The Instruction Processing Cycle  Fetch  Decode  Evaluate Address  Fetch Operands  Execute  Store Result 15 1. Instruction fetch (IF) 2. Instruction decode and register operand fetch (ID/RF) 3. Execute/Evaluate memory address (EX/AG) 4. Memory operand fetch (MEM) 5. Store/writeback result (WB)

16. Remember the Single-Cycle Uarch 16 PCSrc 2 =Br Taken PCSrc 1 =Jump ALU operation bcond Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.] T BW=~(1/T)

17. Dividing Into Stages 17 200ps Is this the correct partitioning? Why not 4 or 6 stages? Why not different boundaries? 100ps200ps 100ps RF write ignore for now Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.]

18. Instruction Pipeline Throughput 18 200 400 600 800 1000 1200 1400 1600 1800 200 400 600 800 1000 1200 1400 800ps 200ps 5-stage speedup is 4, not 5 as predicted by the ideal model. Why?

19. Enabling Pipelined Processing: Pipeline Registers 19 T No resource is used by more than 1 stage! IR D PC F PC D +4PC E +4nPC M AEAE BEBE Imm E Aout M BMBM MDR W Aout W Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.] T/k ps T/k ps

20. Pipelined Operation Example 20 Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.] All instruction classes must follow the same path and timing through the pipeline stages. Any performance impact?

21. Pipelined Operation Example 21 Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.] Is life always this beautiful?

22. Illustrating Pipeline Operation: Operation View 22 MEM EX ID IF Inst 4 WB IF MEM IF MEM EX t0t0 t1t1 t2t2 t3t3 t4t4 t5t5 ID EX IFID IF ID Inst 0 ID IF Inst 1 EX ID IF Inst 2 MEM EX ID IF Inst 3 WB MEM EX WB

23. Illustrating Pipeline Operation: Resource View 23 I0I0 I0I0 I1I1 I0I0 I1I1 I2I2 I0I0 I1I1 I2I2 I3I3 I0I0 I1I1 I2I2 I3I3 I4I4 I1I1 I2I2 I3I3 I4I4 I5I5 I2I2 I3I3 I4I4 I5I5 I6I6 I3I3 I4I4 I5I5 I6I6 I7I7 I4I4 I5I5 I6I6 I7I7 I8I8 I5I5 I6I6 I7I7 I8I8 I9I9 I6I6 I7I7 I8I8 I9I9 I 10 t0t0 t1t1 t2t2 t3t3 t4t4 t5t5 t6t6 t7t7 t8t8 t9t9 t 10 IF ID EX MEM WB

24. Control Points in a Pipeline 24 Identical set of control points as the single-cycle datapath!! Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.]

25. Control Signals in a Pipeline For a given instruction  same control signals as single-cycle, but  control signals required at different cycles, depending on stage  decode once using the same logic as single-cycle and buffer control signals until consumed  or carry relevant “instruction word/field” down the pipeline and decode locally within each or in a previous stage Which one is better? 25

26. Pipelined Control Signals 26 Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.]

27. An Ideal Pipeline Goal: Increase throughput with little increase in cost (hardware cost, in case of instruction processing) Repetition of identical operations  The same operation is repeated on a large number of different inputs Repetition of independent operations  No dependencies between repeated operations Uniformly partitionable suboperations  Processing an be evenly divided into uniform-latency suboperations (that do not share resources) Fitting examples: automobile assembly line, doing laundry  What about the instruction processing “cycle”? 27

28. Instruction Pipeline: Not An Ideal Pipeline Identical operations... NOT!  different instructions do not need all stages - Forcing different instructions to go through the same multi-function pipe  external fragmentation (some pipe stages idle for some instructions) Uniform suboperations... NOT!  difficult to balance the different pipeline stages - Not all pipeline stages do the same amount of work  internal fragmentation (some pipe stages are too fast but all take the same clock cycle time) Independent operations... NOT!  instructions are not independent of each other - Need to detect and resolve inter-instruction dependencies to ensure the pipeline operates correctly  Pipeline is not always moving (it stalls) 28

29. Issues in Pipeline Design Balancing work in pipeline stages  How many stages and what is done in each stage Keeping the pipeline correct, moving, and full in the presence of events that disrupt pipeline flow  Handling dependences Data Control  Handling resource contention  Handling long-latency (multi-cycle) operations Handling exceptions, interrupts Advanced: Improving pipeline throughput  Minimizing stalls 29

30. Causes of Pipeline Stalls Resource contention Dependences (between instructions)  Data  Control Long-latency (multi-cycle) operations 30

31. Dependences and Their Types Also called “dependency” or less desirably “hazard” Dependencies dictate ordering requirements between instructions Two types  Data dependence  Control dependence Resource contention is sometimes called resource dependence  However, this is not fundamental to (dictated by) program semantics, so we will treat it separately 31

32. Handling Resource Contention Happens when instructions in two pipeline stages need the same resource Solution 1: Eliminate the cause of contention  Duplicate the resource or increase its throughput E.g., use separate instruction and data memories (caches) E.g., use multiple ports for memory structures Solution 2: Detect the resource contention and stall one of the contending stages  Which stage do you stall?  Example: What if you had a single read and write port for the register file? 32

33. Data Dependences Types of data dependences  Flow dependence (true data dependence – read after write)  Output dependence (write after write)  Anti dependence (write after read) Which ones cause stalls in a pipelined machine?  For all of them, we need to ensure semantics of the program is correct  Flow dependences always need to be obeyed because they constitute true dependence on a value  Anti and output dependences exist due to limited number of architectural registers They are dependence on a name, not a value We will later see what we can do about them 33

34. Data Dependence Types 34 Flow dependence r 3  r 1 op r 2 Read-after-Write r 5  r 3 op r 4 (RAW) Anti dependence r 3  r 1 op r 2 Write-after-Read r 1  r 4 op r 5 (WAR) Output-dependence r 3  r 1 op r 2 Write-after-Write r 5  r 3 op r 4 (WAW) r 3  r 6 op r 7

35. Pipelined Operation Example 35 Based on original figure from [P&H CO&D, COPYRIGHT 2004 Elsevier. ALL RIGHTS RESERVED.] What if the SUB were dependent on LW?

36. Data Dependence Handling 36

37. Readings for Next Few Lectures P&H Chapter 4.9-4.11 Smith and Sohi, “The Microarchitecture of Superscalar Processors,” Proceedings of the IEEE, 1995  More advanced pipelining  Interrupt and exception handling  Out-of-order and superscalar execution concepts 37

38. How to Handle Data Dependences Anti and output dependences are easier to handle  write to the destination in one stage and in program order Flow dependences are more interesting Five fundamental ways of handling flow dependences  Detect and wait until value is available in register file  Detect and forward/bypass data to dependent instruction  Detect and eliminate the dependence at the software level No need for the hardware to detect dependence  Predict the needed value(s), execute “speculatively”, and verify  Do something else (fine-grained multithreading) No need to detect 38

39. Interlocking Detection of dependence between instructions in a pipelined processor to guarantee correct execution Software based interlocking vs. Hardware based interlocking MIPS acronym? 39

40. Approaches to Dependence Detection (I) Scoreboarding  Each register in register file has a Valid bit associated with it  An instruction that is writing to the register resets the Valid bit  An instruction in Decode stage checks if all its source and destination registers are Valid Yes: No need to stall… No dependence No: Stall the instruction Advantage:  Simple. 1 bit per register Disadvantage:  Need to stall for all types of dependences, not only flow dep. 40

41. Not Stalling on Anti and Output Dependences What changes would you make to the scoreboard to enable this? 41

42. Approaches to Dependence Detection (II) Combinational dependence check logic  Special logic that checks if any instruction in later stages is supposed to write to any source register of the instruction that is being decoded  Yes: stall the instruction/pipeline  No: no need to stall… no flow dependence Advantage:  No need to stall on anti and output dependences Disadvantage:  Logic is more complex than a scoreboard  Logic becomes more complex as we make the pipeline deeper and wider (flash-forward: think superscalar execution) 42

43. Once You Detect the Dependence in Hardware What do you do afterwards? Observation: Dependence between two instructions is detected before the communicated data value becomes available Option 1: Stall the dependent instruction right away Option 2: Stall the dependent instruction only when necessary  data forwarding/bypassing Option 3: … 43

44. Data Forwarding/Bypassing Problem: A consumer (dependent) instruction has to wait in decode stage until the producer instruction writes its value in the register file Goal: We do not want to stall the pipeline unnecessarily Observation: The data value needed by the consumer instruction can be supplied directly from a later stage in the pipeline (instead of only from the register file) Idea: Add additional dependence check logic and data forwarding paths (buses) to supply the producer’s value to the consumer right after the value is available Benefit: Consumer can move in the pipeline until the point the value can be supplied  less stalling 44

45. A Special Case of Data Dependence Control dependence  Data dependence on the Instruction Pointer / Program Counter 45

46. Control Dependence Question: What should the fetch PC be in the next cycle? Answer: The address of the next instruction  All instructions are control dependent on previous ones. Why? If the fetched instruction is a non-control-flow instruction:  Next Fetch PC is the address of the next-sequential instruction  Easy to determine if we know the size of the fetched instruction If the instruction that is fetched is a control-flow instruction:  How do we determine the next Fetch PC? In fact, how do we know whether or not the fetched instruction is a control-flow instruction? 46